It's a good question. Top AIs already exceed the intelligence of about ~1/3rd of humans, at least when it comes to questions that can be verbalized (AIs haven't yet been trained on bricklaying, nor to think ahead to towards future tasks, so almost every human still has an edge over AI somewhere.)
I think a pessimistic (slow) scenario for AI development is that progress levels off for the top models not too far from where we are now, due to energy or data constraints, and the AIs still struggle to break say 110 IQ 5 years from now. AI would still massively transform society, because there are so many unapplied uses of the AI tech already developed, whether it's putting ChatGPT in a robot body and training it on real-world data, or giving them longer time horizons, or giving them more agencies to actually go to Amazon and order something for you, like a real assistant would.
In a fast scenario, which I suspect is slightly less likely, the growth stays on its current trajectory, and in 5 years, many people are asking "why hire a lawyer, when ChatGPT will draft this just as well?"
I'll bet against AI reaching "superintelligence" in 5 years (smarter than almost all humans at tasks, broadly) but of course who can say.
You make me curious to see the offline questions. Though I understand the problem I wonder if there might be a way to cut out scrapers without impacting humans. I have some ideas based on fronting the questions with a service like cloudflare which allows you to redirect based on the requester's source (and indeed to make the bot click on the "I'm not a robot" box, which it ought to fail at). This would let humans see your test questions and exclude the various web crawlers including the AI ones.
It should be possible to test this by putting some other information behind a test block and see if that information shows up in AI programs or indeed in google/bing etc.
If your test takers are "people who read blogs about AI intelligence and like taking IQ tests", it will be amongst smartest 1..5% of humans, so an score of 100 in a test calibrated with them will correspond to ~120..130 with average humans.
That needs to be confirmed with data, but it is possible, though it would also be a bit surprising if Norway Mensa’s norming (which al this is based on) were so off. It could also be in the middle of 100 and what you suggest
Maxim, great work. Where do you see this heading in the near future? Are we going to have AIs that exceed human intelligence in the coming five years?
It's a good question. Top AIs already exceed the intelligence of about ~1/3rd of humans, at least when it comes to questions that can be verbalized (AIs haven't yet been trained on bricklaying, nor to think ahead to towards future tasks, so almost every human still has an edge over AI somewhere.)
I think a pessimistic (slow) scenario for AI development is that progress levels off for the top models not too far from where we are now, due to energy or data constraints, and the AIs still struggle to break say 110 IQ 5 years from now. AI would still massively transform society, because there are so many unapplied uses of the AI tech already developed, whether it's putting ChatGPT in a robot body and training it on real-world data, or giving them longer time horizons, or giving them more agencies to actually go to Amazon and order something for you, like a real assistant would.
In a fast scenario, which I suspect is slightly less likely, the growth stays on its current trajectory, and in 5 years, many people are asking "why hire a lawyer, when ChatGPT will draft this just as well?"
I'll bet against AI reaching "superintelligence" in 5 years (smarter than almost all humans at tasks, broadly) but of course who can say.
You make me curious to see the offline questions. Though I understand the problem I wonder if there might be a way to cut out scrapers without impacting humans. I have some ideas based on fronting the questions with a service like cloudflare which allows you to redirect based on the requester's source (and indeed to make the bot click on the "I'm not a robot" box, which it ought to fail at). This would let humans see your test questions and exclude the various web crawlers including the AI ones.
It should be possible to test this by putting some other information behind a test block and see if that information shows up in AI programs or indeed in google/bing etc.
If your test takers are "people who read blogs about AI intelligence and like taking IQ tests", it will be amongst smartest 1..5% of humans, so an score of 100 in a test calibrated with them will correspond to ~120..130 with average humans.
That needs to be confirmed with data, but it is possible, though it would also be a bit surprising if Norway Mensa’s norming (which al this is based on) were so off. It could also be in the middle of 100 and what you suggest